A Neural Approach to Cursive Handwritten Character Recognition Using Features Extracted from Binarization Technique - Complex System Modelling and Control Through Intelligent Soft Computations - page 746

Fig. 14 Architecture of the neural network used in the recognition system

The weight update for the output neurons of the network can be determined as:
The error signal at the output of neuron j at iteration n is:

e j ð n Þ¼ d j ð n Þ y j ð n Þ ð 1 Þ

The instantaneous value of error for neuron j is 2 e j ð

n Þ .

eð Þ

of total error is obtained by summing 2 e j ð

Þ

The instantaneous value n n of all

neurons in output layer

2 X
j 2 c

1

e j ð

eð Þ¼ Þ ð Þ

n n 2

where, includes all neurons in output layer.

Average squared error is given by

' c '

N X

N
n ¼ 1 eð

1

e avg ¼ Þ ð Þ

n 3

e avg should be mini-
mized. Back-propagation is used to update the weights. Induced local

where, N is the total number of patterns in training set and

field v j ð n Þ

produced at input of activation function is given by

X

m

v j ð n Þ¼ w ji ð n Þ y i ð n Þ ð 4 Þ

i ¼ 0

where ' m ' is the number of inputs applied to neuron ' j ' , so the output can be written

as:

y j ð n Þ¼/ j ð v j ð n ÞÞ ð 5 Þ

.